I run a couple of small open models on my own computer for personal projects, mostly because I like knowing where my data goes. Until this year I wouldn’t have suggested a business do the same with real work. The models that fitted on a desk were toys next to the ones in the cloud, and the gap was too wide to argue about.

Six months on, late April looks like the turning point. Poolside released Laguna XS.2, a free coding model that the company says will run on a Mac with 36GB of memory. On SWE-bench Verified, a test built from real software bugs, it fixes 69.9% of them, against 73.3% for Anthropic’s Claude Haiku 4.5, which you can only rent. Alibaba had released its latest mid-sized Qwen model earlier that month, and Nvidia’s Nemotron 3 Nano Omni, which handles text, images, audio and video on a single workstation, came out the same day as Laguna. For most of the internal work small businesses hand to AI, an open model on hardware you control is now good enough.

Where the work runs is now a choice

Most of us have spent three years treating AI as something we rent, and that made sense while every good model lived in someone else’s data centre. The design question now is where each job should run, rather than which cloud or mix of clouds to buy. In March I argued that most AI work doesn’t need a frontier model, and the same logic goes a step further. If a task doesn’t need the best model, it may be able to stay in the building too.

Start by listing the AI jobs you run today and the ones you’re planning. Then ask three questions of each. Does the data have to leave your network? Does the task need the strongest model available, or would a good one do? Does it need capacity that rises and falls with demand? If the answer to all three is no, it’s a candidate for running locally.

Take an accountancy practice summarising client files. The data shouldn’t leave, a good model does the job, and the volume is much the same every week, so it passes on every count. Searching your own documents, drafting first versions and helping a developer with routine code usually pass too. Deep research, long multi-step planning and anything customer-facing at scale usually don’t, because they still benefit from frontier models and from capacity you only pay for when you use it.

The data question has a governance side. When I wrote about DeepSeek’s prices in May, the sticking point I raised was that using its API means sending your data to a Chinese provider. Running open weights on your own machine removes that worry, although you still have to be comfortable with who trained the model.

Where it tips back to the cloud

Local isn’t free, and for plenty of businesses the cloud will stay the sensible answer. Nvidia’s DGX Spark, the desktop box it sells for this kind of work, launched at $3,999 last October. It went up to $4,699 in February because of memory shortages, and TechnoEdge reports it’s now listed at $6,950 in the US.

Spread that over three years and it’s a little under $200 a month, before electricity and before anyone’s time looking after it. At Claude Haiku 4.5’s published price of $1 per million tokens read and $5 per million written, $200 a month buys roughly 100 million tokens of reading and 20 million of writing. That’s a great deal of document summarising for a small firm. Below that volume the cloud is cheaper, and DeepSeek’s V4 Flash, at 14 cents per million input tokens, moves the bar higher still if its output is good enough for the job.

So my rough rule is that local pays for itself only when a workload is both large and steady from week to week. Spiky or occasional work belongs in the cloud. I’m not sure how quickly that threshold will move, since hardware is getting dearer while API prices keep falling. What I am sure of is that for most small businesses the data is a stronger reason to go local than the money.

That matters if you’re about to renew an annual AI budget or sign up for reserved cloud capacity. Sort your workloads first, and commit cloud spend only to the ones that need it.

The cloud still has plenty of work to do, and the strongest models will live there for some time yet. For much of what a small business runs through AI today, though, the test will keep more of it on your own hardware than most of us have been assuming, and whatever stays in the cloud will be there for a reason you can name.